AI Core Concepts, Visually Explained

The landscape

Large language models — a kind of generative machine learning — have become our front door to the wider landscape of “AI.”

What Is an LLM

At its simplest, a large language model is a function that turns a prompt into a statistically likely response. Under the hood, it reads the prompt as a series of tokens (words and word-pieces), runs them through its billions of parameters (or weights) — the numeric dials below, set during training — via a vast chain of matrix multiplications, and generates a new sequence of tokens, one at a time, that it returns as the response. Running a trained model this way is called inference.

Basic App Stack

On its own, an LLM is just a function that turns a prompt into text — useful, but not yet a product. App developers make it usable by packaging it: app logic on top and compute underneath, both of which the model needs to reach real users. App logic is the layer the developer builds — UI, accounts and billing, the system prompt, memory. Compute is the hardware that runs the model. Claude and ChatGPT are one such packaging, but these three layers describe any AI app: a company building its own assistant picks its own model and its own compute, then writes the app logic around them.

More Sophisticated Apps

Developers extend an App’s functionality two ways: by embedding multiple models inside it, or by connecting it outward to other AI apps, data sources, and software. To do both, they introduce a harness in the App Logic — a coordinator that shares context between the model calls and the connected apps.

Agents

Agents are AI apps designed to pursue a goal and take actions within a specific domain.

The new concept is the prompt: in a normal app, a prompt can be any request; in an agent, the prompt becomes a goal that App Logic pursues by looping through model calls, software actions, and observations until the work is done.

For example, an agent can turn “RSVP yes to the invite from my friend” into a short workflow: understand the goal, inspect the invite, send the RSVP through software, and report back.

The model harness

One way to use these ideas: a harness coordinates several specialized model calls — plan, execute, judge — and loops until the output meets the bar.

Where it runs

Any model runs in one of two homes — cloud-hosted or self-hosted — and the choice decides who sees your data.

Inference providers

If you go cloud, three kinds of provider can serve the model: the model lab itself, a cloud platform, or a dedicated inference provider.